Improved performance of error correcting output codes for multiclass classification

Abstract

Some of the best performing classification algorithms to date are only capable of dealing with binary problems, or else require complex adaptations to make them suitable for multi-class situations. A common approach to bypass this issue is to restructure the multi-class problem into several binary ones, then use ensemble techniques to solve each of the binary problems and formulate a single classification result. Some of the most widely known ensemble techniques are the one-vs-one and the one-vs-all. Over recent years, Error-Correcting Output Codes (ECOC) have been devised to improve on this concept by adding a layer of error-correction techniques to the process. Although this concept is still relatively new, research publications show that it has a lot of potential. This dissertation sets out to identify the current state-of-the-art ECOC algorithm with the intention of improving on it. The loss-weighted algorithm published by Escalera et al matched the necessary criteria and was selected as a benchmark algorithm for this project. Whilst implementing it to replicate the results published by the authors, a limitation was noted in the way the algorithm disregards information generated by dichotomies whose input is assumed to belong to a class that was not used during training. Based on this, a new algorithm has been proposed, implemented and tested to establish if classification performance would improve if the data ignored by the loss-weighted algorithm had to be considered. Results show that in general, the proposed ECOC decoding algorithm outperforms the current state-of-the-art counterpart by an average of 2.1%, given that the classification problem meets two specific criteria.